Ship Classification in High-Resolution SAR Images Based on CNN Regional Feature Fusion

Abstract

Ship classification is a multi-level and multi-category target recognition research, among them, military-civilian ship classification plays a significance role in maritime target monitoring. Nowadays, with the development of spaceborne SAR systems, the resolution and quality of SAR images have further been improved. Ship targets in high-resolution SAR images show clear geometric structure and scattering characteristics, which is conductive to military-civilian ship classification. Considering the adaptive feature extraction ability of convolutional neural network (CNN) and the structural characteristics of ships in high-resolution SAR images, a military-civilian ship classification method based on CNN regional feature fusion is proposed in this paper. Proposed method can be divided into three parts: obtaining Minimum Bounding Rectangle (MBR) of ship target in SAR images, global and local CNN feature extraction, and adaptive feature fusion for ship classification. To verify method proposed in this paper, we conducted extensive experiments on the high-resolution SAR images obtained by GF-3 satellite. The experimental results show that our method can effectively classify military and civilian ships with an accuracy rate of over 90%.